Test Sample Accuracy Scales with Training Sample Density in Neural Networks
This work addresses improving prediction reliability for neural networks, particularly on out-of-distribution samples, but is incremental as it builds on existing density-based analysis.
The study tackled the problem of predicting neural network test accuracy based on training sample density in representation space, finding that ranking and discarding test samples with high error bounds can increase prediction accuracy by up to 20% on image classification datasets.
Intuitively, one would expect accuracy of a trained neural network's prediction on test samples to correlate with how densely the samples are surrounded by seen training samples in representation space. We find that a bound on empirical training error smoothed across linear activation regions scales inversely with training sample density in representation space. Empirically, we verify this bound is a strong predictor of the inaccuracy of the network's prediction on test samples. For unseen test sets, including those with out-of-distribution samples, ranking test samples by their local region's error bound and discarding samples with the highest bounds raises prediction accuracy by up to 20% in absolute terms for image classification datasets, on average over thresholds.